Quick answer
Do not treat Agent Executor as generally available or stable. Google’s May 20, 2026 announcement calls it Preview. The current google/ax repository warns that its core, resumption protocols, and runtime specifications are being refined and will introduce major breaking changes before a stable release. Its Kubernetes guide adds that the manifests, scripts, and runtime behavior are experimental, incomplete, and may break without notice.
A team can still evaluate AX, but the safe unit is a pinned revision, a reversible workload, an isolated cohort, and explicit stop conditions—not “Agent Executor” as an evergreen product promise.
Preview evidence map
| Signal | What the source says | Decision consequence |
|---|---|---|
| Launch lifecycle | Preview | No GA or stable-support inference |
| Repository lifecycle | Active early development | Expect interface and behavior churn |
| Breaking changes | Major changes before stable release | Pin, test, migrate, and retain exit path |
| External pull requests | Temporarily paused during architecture stabilization | Do not rely on normal contribution flow for urgent fixes |
| Kubernetes path | Experimental and incomplete | Separate lab success from production approval |
| Product form | Self-hosted, not managed | Your team owns operations and most failure recovery |
| Framework role | Harness-agnostic runtime, not an agent framework | Agent quality and tool safety remain outside AX |
These are current project statements, not criticisms. They are precisely the facts a platform review should preserve.
A production-readiness scorecard
Compatibility
Pin the repository revision, CLI and gRPC protocol, Go dependencies, harness packages, Agent Substrate commit, CRDs, images, configuration schema, event-log schema, and snapshot format. Prove forward migration and rollback using unfinished conversations, not only empty environments.
Reliability
Measure recovery point and recovery time for controller failure, database failure, actor eviction, snapshot failure, model timeout, tool ambiguity, and client disconnect. Verify duplicate prevention for external side effects.
Security
Prove caller identity, resource authorization, actor and tenant isolation, secret scope, network policy, image provenance, dependency scanning, audit access, and deletion. A “secure-by-design sandbox” statement does not replace adversarial evidence.
Operations
Define SLOs, capacity limits, on-call ownership, telemetry, redaction, runbooks, database and bucket backups, cost budgets, kill switches, incident escalation, and a tested uninstall or migration plan.
Support and governance
Identify who accepts preview change risk and who owns upstream issue escalation. Document license review, privacy and retention, model and tool vendors, regional data paths, and the business approver for the first workload.
Use a bounded pilot contract
A defensible pilot states:
- one pinned AX revision and dependency set;
- one low-consequence, reversible task;
- a small named user or traffic cohort;
- read-only tools before write tools;
- maximum event, snapshot, model, and infrastructure spend;
- explicit reliability and isolation acceptance tests;
- a freeze condition for upstream changes;
- rollback and data-deletion evidence;
- an end date and named go/no-go owner.
The pilot should stop if state cannot be migrated, a duplicate side effect cannot be reconciled, tenant isolation fails, a critical dependency must float, or required observability is absent.
Use the AX runtime guide for system boundaries, the Kubernetes deployment guide for version coupling, and the local readiness checklist for the release decision.
Frequently asked questions
Is Agent Executor generally available?
No. The launch label is Preview, and the current repository warns of active early development.
Can a team deploy it in production?
Only as its own evidence-backed risk decision. Upstream availability does not establish your workload’s security, reliability, support, or compatibility.
Are breaking changes only a theoretical risk?
No. The repository directly says major breaking changes are expected before stable release, and the deployment guide warns of unannounced breakage.
Official sources
- Google Cloud: Agent Executor announcement
- Google: AX repository
- Google: experimental Kubernetes deployment guide
Source check: August 12, 2026; AX revision 2bcc1637b3c106c16963c5c5464aedb46c6da031. Recheck lifecycle, warnings, contribution policy, protocols, roadmap, and deployment status before every release decision.